Embodiments are directed to techniques for allowing a data storage system to be able to flush data to underlying storage when the bandwidth is high without excessively impacting the maximum latency. This may be accomplished by utilizing asynchronous flushing and by throttling incoming writes by preventing too many asynchronous flushes from happening when the amount of available cache space is too small. In addition, an improved system employing Copy on First Write (CoFW) may initiate write throttling only once the amount of available write cache drops below a dynamically-calculated threshold that accounts for an amount of space actually needed to store CoFW buffers. In addition, only a portion of the write caching process is placed under the protection of a mutex or a spinlock, allowing a significant portion of the write caching for any given write operation to be performed without needing the mutex or spinlock, allowing some increased parallelization.